<div dir="ltr">This patch fixes the issue. I haven't checked our lit tests yet to see what effect this has and if we can commit it.<div><font color="#000000"><br></font></div><div>





<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000">lib/CodeGen/SelectionDAG/DAGCombiner.cpp</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><font style="background-color:rgb(255,255,255)" color="#000000"><span class="gmail-s2" style="font-variant-ligatures:no-common-ligatures">@@ -2967,7 +2967,8 @@</span><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures"> SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn,</span></font></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space">   </span>// Iff the flag result is dead:</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space">   </span>// (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry)</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space">   </span>if ((N0.getOpcode() == ISD::ADD ||</font></span></p>
<p class="gmail-p2" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000">- <span class="gmail-Apple-converted-space">      </span>(N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0)) &&</font></span></p>
<p class="gmail-p3" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000">+ <span class="gmail-Apple-converted-space">      </span>(N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0 &&</font></span></p>
<p class="gmail-p3" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000">+<span class="gmail-Apple-converted-space">        </span>N0.getValue(1) != CarryIn)) &&</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space">       </span>isNullConstant(N1) && !N->hasAnyUseOfValue(1))</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space">     </span>return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(),</font></span></p>
<p class="gmail-p1" style="margin:0px;font-variant-numeric:normal;font-variant-east-asian:normal;font-stretch:normal;font-size:11px;line-height:normal;font-family:Menlo"><span class="gmail-s1" style="font-variant-ligatures:no-common-ligatures;background-color:rgb(255,255,255)"><font color="#000000"><span class="gmail-Apple-converted-space" style="">                        </span>N0.getOperand(0), N0.getOperand(1), CarryIn);</font></span></p></div><div><br clear="all"><div><div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature">~Craig</div></div><br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, Jul 3, 2019 at 9:40 AM Marc <<a href="mailto:mh@mh-sec.de">mh@mh-sec.de</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Hi Craig,<br>
<br>
On 03.07.19 17:33, Craig Topper wrote:<br>
> Don't the CreateICmp calls return a Value* with an i1 type? But then<br>
> they are added to an i8 type? Not sure that works. <br>
<br>
I had that initially:<br>
 auto cf = IRB.CreateICmpULT(Incr, ConstantInt::get(Int8Ty, 1));<br>
 auto carry = IRB.CreateZExt(cf, Int8Ty);<br>
 Incr = IRB.CreateAdd(Incr, carry);<br>
it makes no difference to the generated assembly<br>
<br>
<br>
> Have you tried using the llvm.uadd.with.overflow.i8 intrinsic?<br>
<br>
we have tried this:<br>
<br>
 CallInst *AddOv =<br>
IRB.CreateBinaryIntrinsic(Intrinsic::uadd_with_overflow, Counter,<br>
ConstantInt::get(Int8Ty, 1));<br>
 AddOv->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
 Value *SumWithOverflowBit = AddOv;<br>
 Incr = IRB.CreateAdd(IRB.CreateExtractValue(SumWithOverflowBit, 0),<br>
 IRB.CreateZExt(IRB.CreateExtractValue(SumWithOverflowBit, 1), Int8Ty));<br>
<br>
but that generated exactly the same as<br>
<br>
 auto cf = IRB.CreateICmpULT(Incr, ConstantInt::get(Int8Ty, 1));<br>
 Incr = IRB.CreateAdd(Incr, cf);<br>
<br>
<br>
All the above create (with llvm 6.0, 7 and 8):<br>
<br>
 mov    dl,BYTE PTR [rsi+rdi*1]<br>
 mov    ecx,edx // <- not needed<br>
 add    cl,0x1  // <- should be done to dl instead<br>
 adc    dl,0x1<br>
 mov    BYTE PTR [rsi+rdi*1],dl<br>
<br>
<br>
Thanks!<br>
<br>
Regards,<br>
Marc<br>
<br>
> <br>
> ~Craig<br>
> <br>
> <br>
> On Wed, Jul 3, 2019 at 4:01 AM Marc via llvm-dev<br>
> <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a> <mailto:<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>>> wrote:<br>
> <br>
>     Hello,<br>
> <br>
>     I have an optimisation issue in an llvm IR pass - the issue being that<br>
>     unnecessary instructions are generated in the final assembly (with -O3).<br>
> <br>
>     I want to create the following assembly snippet:<br>
> <br>
>       mov    dl,BYTE PTR [rsi+rdi*1]<br>
>       add    dl,0x1<br>
>       adc    dl,0x0<br>
>       mov    BYTE PTR [rsi+rdi*1],dl<br>
> <br>
>     however what is created is (variant #1):<br>
> <br>
>       mov    dl,BYTE PTR [rsi+rdx*1]<br>
>       add    dl,0x1<br>
>       cmp    dl,0x1 // <- not needed<br>
>       adc    dl,0x0<br>
>       mov    BYTE PTR [rsi+rdi*1],dl<br>
> <br>
>     in variant #2 the following is created:<br>
> <br>
>      mov    dl,BYTE PTR [rsi+rdi*1]<br>
>      mov    ecx,edx // <- not needed<br>
>      add    cl,0x1  // <- should be done to dl instead<br>
>      adc    dl,0x1<br>
>      mov    BYTE PTR [rsi+rdi*1],dl<br>
> <br>
>     Far below are both variants with the full code around it, however the<br>
>     difference in both variants is this:<br>
> <br>
>     //variant1<br>
>     auto cf = IRB.CreateICmpEQ(Incr, ConstantInt::get(Int8Ty, 0));<br>
>     Incr = IRB.CreateAdd(Incr, cf);<br>
> <br>
>     //variant2<br>
>     auto cf = IRB.CreateICmpULT(Incr, ConstantInt::get(Int8Ty, 1));<br>
>     Incr = IRB.CreateAdd(Incr, cf);<br>
> <br>
>     //interestingly this totally different approach creates the same<br>
>     instructions as variant2<br>
>     CallInst *AddOv = IRB.CreateBinaryIntrinsic(Intrinsic::uadd_with_over<br>
>     flow, Counter, ConstantInt::get(Int8Ty, 1));<br>
>     AddOv->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
>     Value *SumWithOverflowBit = AddOv;<br>
>     Incr = IRB.CreateAdd(IRB.CreateExtractValue(SumWithOverflowBit, 0),<br>
>     IRB.CreateZExt(IRB.CreateExtractValue(SumWithOverflowBit, 1), Int8Ty));<br>
> <br>
>     So - How do I have to write the code that the target code has a chance<br>
>     of being generated?<br>
>     For me its the same result on LLVM 6.0 and 7.<br>
> <br>
>     Alternatively<br>
>       add    BYTE PTR [rsi+rdi*1],0x1<br>
>       adc    BYTE PTR [rsi+rdi*1],0x0<br>
>      would work as well.<br>
> <br>
>     Thank you very much!<br>
> <br>
>     Regards,<br>
>     Marc<br>
> <br>
>     -> Below now is the full llvm IR pass code<br>
> <br>
>     // vvv same code before both variants<br>
>     LoadInst *PrevLoc = IRB.CreateLoad(AFLPrevLoc);<br>
>     PrevLoc->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
>     Value *PrevLocCasted = IRB.CreateZExt(PrevLoc, IRB.getInt32Ty());<br>
> <br>
>      LoadInst *MapPtr = IRB.CreateLoad(AFLMapPtr);<br>
>     MapPtr->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
>     Value *MapPtrIdx = IRB.CreateGEP(MapPtr, IRB.CreateXor(PrevLocCasted,<br>
>     CurLoc));<br>
> <br>
>     LoadInst *Counter = IRB.CreateLoad(MapPtrIdx);<br>
>     Counter->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
> <br>
>     Value *Incr = IRB.CreateAdd(Counter, ConstantInt::get(Int8Ty, 1));<br>
>     // ^^^ same code before both variants<br>
> <br>
>     // CODE FOR VARIANT #1<br>
>     auto cf = IRB.CreateICmpEQ(Incr, ConstantInt::get(Int8Ty, 0));<br>
>     Incr = IRB.CreateAdd(Incr, cf);<br>
> <br>
>     // CODE FOR VARIANT #2<br>
>     auto cf = IRB.CreateICmpULT(Incr, ConstantInt::get(Int8Ty, 1));<br>
>     Incr = IRB.CreateAdd(Incr, cf);<br>
> <br>
>     // vvv same code after both variants<br>
>     IRB.CreateStore(Incr,<br>
>     MapPtrIdx)->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C,<br>
>     None));<br>
>     StoreInst *Store = IRB.CreateStore(ConstantInt::get(Int32Ty, cur_loc >><br>
>     1), AFLPrevLoc);<br>
>     Store->setMetadata(M.getMDKindID("nosanitize"), MDNode::get(C, None));<br>
>     // ^^^ same code after both variants<br>
> <br>
> <br>
>     -- <br>
>     Marc Heuse<br>
>     <a href="http://www.mh-sec.de" rel="noreferrer" target="_blank">www.mh-sec.de</a> <<a href="http://www.mh-sec.de" rel="noreferrer" target="_blank">http://www.mh-sec.de</a>><br>
> <br>
>     PGP: AF3D 1D4C D810 F0BB 977D  3807 C7EE D0A0 6BE9 F573<br>
>     _______________________________________________<br>
>     LLVM Developers mailing list<br>
>     <a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a> <mailto:<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>><br>
>     <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
> <br>
<br>
-- <br>
Marc Heuse<br>
Mobil: +49 177 9611560<br>
Fax: +49 30 37309726<br>
<a href="http://www.mh-sec.de" rel="noreferrer" target="_blank">www.mh-sec.de</a><br>
<br>
Marc Heuse - IT-Security Consulting<br>
Winsstr. 68<br>
10405 Berlin<br>
<br>
Ust.-Ident.-Nr.: DE244222388<br>
PGP: AF3D 1D4C D810 F0BB 977D  3807 C7EE D0A0 6BE9 F573<br>
</blockquote></div>